{"about":{"site":"https://codewithpapers.app","non_affiliation":"Code with Papers and Syntology are not affiliated with, endorsed by, or sponsored by Papers with Code, Meta, or the pwc-archive mirror.","licence":"CC BY-SA 4.0","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","attribution":"https://codewithpapers.app/attribution","modified":"archive material modified by Syntology; see the attribution page"},"url":"/paper/visualisation-and-diagnostic-classifiers","title":"Visualisation and 'diagnostic classifiers' reveal how recurrent and recursive neural networks process hierarchical structure","arxiv_id":"1711.10203","date":"2017-11-28","proceeding":null,"authors":["Dieuwke Hupkes","Sara Veldhoen","Willem Zuidema"],"abstract":"We investigate how neural networks can learn and process languages with\nhierarchical, compositional semantics. To this end, we define the artificial\ntask of processing nested arithmetic expressions, and study whether different\ntypes of neural networks can learn to compute their meaning. We find that\nrecursive neural networks can find a generalising solution to this problem, and\nwe visualise this solution by breaking it up in three steps: project, sum and\nsquash. As a next step, we investigate recurrent neural networks, and show that\na gated recurrent unit, that processes its input incrementally, also performs\nvery well on this task. To develop an understanding of what the recurrent\nnetwork encodes, visualisation techniques alone do not suffice. Therefore, we\ndevelop an approach where we formulate and test multiple hypotheses on the\ninformation encoded and processed by the network. For each hypothesis, we\nderive predictions about features of the hidden state representations at each\ntime step, and train 'diagnostic classifiers' to test those predictions. Our\nresults indicate that the networks follow a strategy similar to our\nhypothesised 'cumulative strategy', which explains the high accuracy of the\nnetwork on novel expressions, the generalisation to longer expressions than\nseen in training, and the mild deterioration with increasing length. This is\nturn shows that diagnostic classifiers can be a useful technique for opening up\nthe black box of neural networks. We argue that diagnostic classification,\nunlike most visualisation techniques, does scale up from small networks in a\ntoy domain, to larger and deeper recurrent networks dealing with real-life\ndata, and may therefore contribute to a better understanding of the internal\ndynamics of current state-of-the-art models in natural language processing.","url_abs":"http://arxiv.org/abs/1711.10203v2","url_pdf":"http://arxiv.org/pdf/1711.10203v2.pdf","source":{"archive":"pwc-archive (Hugging Face), CC BY-SA 4.0","snapshot":"2025-07-28","licence_url":"https://creativecommons.org/licenses/by-sa/4.0/legalcode","row_kind":"abstracts"},"code_links":[{"paper_slug":"visualisation-and-diagnostic-classifiers","repo_url":"https://github.com/dieuwkehupkes/processing_arithmetics","is_official":1,"mentioned_in_paper":1,"mentioned_in_github":0,"framework":"none","reach":{"status":"unanswered"}}],"tasks":[{"task_slug":"diagnostic","task_name":"Diagnostic"}],"methods":[],"datasets_introduced":[],"methods_introduced":[],"results":[],"syntology":{"atlas_url":"https://app.syntology.ai/?focus=1711.10203","mcp":null,"developers":"https://syntology.ai/developers"},"arxiv_metadata":null,"syntology_extracted_results":null}